salisai/urdu-tts-mini

Dataset Card for Urdu-TTS-Mini

A curated Urdu speech dataset for Text-to-Speech (TTS) and Automatic Speech Recognition (ASR) research. Audio segments are extracted from publicly available YouTube speech content, processed through a multi-stage quality pipeline, and annotated with Urdu transcriptions. This is a mini release intended to validate the preprocessing pipeline and establish a quality baseline for future large-scale versions.

Dataset Details

Dataset Description

Urdu-TTS-Mini is a small-scale, high-quality Urdu speech dataset containing 5,360 short audio clips totaling 3.74 hours. It is designed for training and evaluating speech models in a low-resource language setting. The dataset was built using a fully automated pipeline covering audio extraction, segmentation, language filtering, quality scoring, and transcript alignment.

• —Curated by: Sayed Ali Shah
• —Shared by [optional]: Sayed Ali Shah
• —Language(s) (NLP): Urdu (ur)
• —License: CC BY 4.0

Dataset Sources [optional]

• —Repository: https://huggingface.co/datasets/salisai/urdu-tts-mini

Uses

Direct Use

This dataset is suitable for:

• —Training Urdu Text-to-Speech (TTS) systems (e.g., FastSpeech, VITS)
• —Fine-tuning Automatic Speech Recognition (ASR) models (e.g., Whisper)
• —Urdu speech representation learning
• —Benchmarking small-scale Urdu speech models

Out-of-Scope Use

This dataset is not suitable for:

• —Production-grade TTS systems requiring broad speaker diversity
• —Speaker identification or verification tasks, as speaker metadata is not provided
• —Applications requiring demographically balanced speech data
• —Any use that violates the terms of the original YouTube source material

Dataset Structure

The dataset consists of short WAV audio clips and a metadata CSV file.

urdu-tts-mini/
├── audio/
│   ├── ur_0000001.wav
│   ├── ur_0000002.wav
│   └── ...
└── metadata.csv

Summary statistics:

Dataset Creation

Curation Rationale

Urdu is a low-resource language in speech AI. Despite being spoken by over 230 million people, high-quality, openly available Urdu speech datasets remain scarce. This dataset is part of an ongoing effort to build clean, structured Urdu speech resources that enable training of modern speech models.

Source Data

Data Collection and Processing

Audio was sourced from publicly available YouTube content featuring Urdu speech. The processing pipeline consisted of the following stages:

1. 1.Audio extraction from long-form speech content
2. 2.Forced segmentation into short utterances
3. 3.Urdu language detection and filtering
4. 4.Audio normalization and resampling to 22,050 Hz
5. 5.Silence removal
6. 6.SNR-based quality filtering (clips below threshold discarded)
7. 7.Transcript cleaning and alignment

Clips shorter than 1.5 seconds or longer than 15 seconds were excluded. Only segments with confirmed Urdu transcripts and sufficient signal quality were retained.

Who are the source data producers?

The source audio originates from YouTube creators producing Urdu-language content. The dataset does not include speaker-level metadata, and no personally identifiable information was deliberately collected.

Annotations [optional]

Annotation process

Transcripts were generated through automated speech recognition and alignment, followed by rule-based cleaning to remove non-Urdu text segments and normalize script formatting.

Who are the annotators?

Annotations were produced automatically as part of the processing pipeline. No human annotators were involved in this release.

Personal and Sensitive Information

The dataset is derived from publicly available YouTube content. No effort was made to collect personally identifiable information. Speaker identities are not tracked or disclosed. Users should nonetheless exercise caution and ensure compliance with the original content licenses.

Bias, Risks, and Limitations

• —Source bias: All audio is sourced from YouTube, which skews toward specific speaking styles, recording environments, and content types.
• —Speaker imbalance: The dataset is not balanced across speakers, genders, ages, or regional dialects of Urdu.
• —Accent variation: YouTube content may include speakers with non-native accents or code-switching between Urdu and other languages.
• —Scale: At 5,360 clips, this is a proof-of-concept release and is not representative of the full diversity of Urdu speech.
• —Transcript quality: Transcripts are automatically generated and may contain errors, particularly for fast speech or overlapping audio.

Recommendations

Users should treat this dataset as a starting point rather than a production-ready resource. Models trained solely on this data may not generalize well across speakers or dialects. It is recommended to supplement with additional Urdu speech data where possible, and to validate transcript quality before use in high-accuracy applications.

Citation [optional]

BibTeX:

@dataset{urdu_tts_mini,
  author    = {Sayed Ali Shah},
  title     = {urdu-tts-mini},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/salisai/urdu-tts-mini}
}

APA:

Shah, S. A. (2025). urdu-tts-mini [Dataset]. Hugging Face. https://huggingface.co/datasets/salisai/urdu-tts-mini

Glossary [optional]

• —TTS (Text-to-Speech): The task of synthesizing natural-sounding speech from written text.
• —ASR (Automatic Speech Recognition): The task of transcribing spoken audio into text.
• —SNR (Signal-to-Noise Ratio): A measure of audio quality expressed in decibels (dB). Higher values indicate cleaner audio with less background noise.
• —Low-resource language: A language for which limited annotated data exists for training machine learning models.

More Information [optional]

This dataset is part of a planned series of progressively larger Urdu speech releases:

Dataset Card Authors [optional]

Sayed Ali Shah

Dataset Card Contact

For questions or feedback, please open a discussion on the dataset repository page.